| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 5.63 MB | Adobe PDF |
Authors
Abstract(s)
O paradigma da utilização de novas tecnologias na agricultura e na indústria alimentar
conduz a um aumento da produção e a uma potencial diminuição das perdas ao longo da
cadeia de abastecimento. Porém, novos desafios surgem na modelação desta indústria,
de modo a criar ambientes fiáveis com estas novas ferramentas.
Esta dissertação tem como objetivo estudar e aplicar duas dessas novas tecnologias, a
visão computacional em conjunto com a robótica colaborativa, de modo a criar um
sistema não destrutivo, ou seja, sem danificar o fruto, de fácil implementação e
economicamente favorável para atuação em linhas de processamento de frutos e
vegetais, auxiliando de maneira direta na deteção de falhas e controlo de qualidade.
Foi desenvolvido um sistema utilizando um Raspberry Pi 5 para captura de imagens dos
frutos por meio de uma PiCamera módulo 3. As imagens foram enviadas a um braço
robótico UR3e da Universal Robots via cabo Ethernet utilizando um código Python que
integra funções desenvolvidas pela própria empresa e funções próprias desenvolvidas
especificamente para esta dissertação.
Foram desenvolvidos quatro modelos de deteção de objetos utilizando o TensorFlow
Object Detection API, convertendo-os posteriormente em TensorFlow Lite para detetar
dois tipos de fruta (laranja e tomate) fazendo uso de técnicas de aprendizagem profunda.
Cada fruto teve dois tipos de modelo, com variação de classes no modelo de tomate,
porém com a mesma base de dados e com duas bases de imagens e número de classes
diferentes para o modelo de laranja, o de laranja com 2 classes obteve 69,85% mAP,
laranja 4 classes 46,37% mAP, tomate quatro classe 78,69% mAP e tomate duas classes
81,43% mAP para uma IoU de 0,5.
Por fim, criou-se uma área de trabalho retangular fiável para atuação do braço robótico
em conjunto com a visão computacional. Após a realização de 640 testes de manipulação,
obteve-se uma área fiável de 262 x 250 mm.
The paradigm of using new technologies in agriculture and the food industry is leading to an increase in production and a potential reduction in losses along the supply chain. However, new challenges arise in modelling this industry in order to create reliable environments with these new tools. This dissertation aims to study and apply two of these new technologies, computer vision in conjunction with collaborative robotics, in order to create a non-destructive system, i.e. without damaging the fruit, that is easy to implement and economically favourable for use in fruit and vegetable processing lines, directly helping with fault detection and quality control. A system was developed using a Raspberry Pi 5 to capture images of the fruit using a PiCamera module 3. The images were sent to a Universal Robots UR3e robotic arm via Ethernet cable using Python code that integrates functions developed by the company itself and its own functions developed specifically for this dissertation. Four object detection models were developed using the TensorFlow Object Detection API and then converted into TensorFlow Lite to detect two types of fruit (oranges and tomatoes) using deep learning techniques. Each fruit had two types of model, with a variation of classes in the tomato model, but with the same database and with two different image bases and number of classes for the orange model, the orange with 2 classes obtained 69.85% mAP, orange 4 classes 46.37% mAP, tomato four classes 78.69% mAP and tomato two classes 81.43% mAP for an IoU of 0.5. Finally, a reliable rectangular work area was created for the robotic arm to operate in conjunction with computer vision. After carrying out 640 manipulation tests, a reliable area of 262 x 250 mm was obtained.
The paradigm of using new technologies in agriculture and the food industry is leading to an increase in production and a potential reduction in losses along the supply chain. However, new challenges arise in modelling this industry in order to create reliable environments with these new tools. This dissertation aims to study and apply two of these new technologies, computer vision in conjunction with collaborative robotics, in order to create a non-destructive system, i.e. without damaging the fruit, that is easy to implement and economically favourable for use in fruit and vegetable processing lines, directly helping with fault detection and quality control. A system was developed using a Raspberry Pi 5 to capture images of the fruit using a PiCamera module 3. The images were sent to a Universal Robots UR3e robotic arm via Ethernet cable using Python code that integrates functions developed by the company itself and its own functions developed specifically for this dissertation. Four object detection models were developed using the TensorFlow Object Detection API and then converted into TensorFlow Lite to detect two types of fruit (oranges and tomatoes) using deep learning techniques. Each fruit had two types of model, with a variation of classes in the tomato model, but with the same database and with two different image bases and number of classes for the orange model, the orange with 2 classes obtained 69.85% mAP, orange 4 classes 46.37% mAP, tomato four classes 78.69% mAP and tomato two classes 81.43% mAP for an IoU of 0.5. Finally, a reliable rectangular work area was created for the robotic arm to operate in conjunction with computer vision. After carrying out 640 manipulation tests, a reliable area of 262 x 250 mm was obtained.
Description
Keywords
Controlo de Qualidade Raspberry Pi Robotica Ur3e Visão Computacional